فهرست نشریات علمی دسترسی آزاد ایران

This paper conducts a comparative evaluation of two advanced large language models (LLMs) — ChatGPT-4 and DeepSeek v3—utilizing 80 algorithmic problems from Code forces categorized into four difficulty levels: Easy (800–1100), Intermediate (1200–1600), Advanced (1700–2000), and Expert (2100–2400), focusing on code generation in Python. Standardized prompts and controlled testing conditions enable the assessment of models on accuracy, effi-ciency, and code readability. As the complexity of issues increases, DeepSeek frequently out-performs ChatGPT in both accuracy and efficiency, despite both models excelling in simpler tasks. This, however, results in reduced code clarity and increased memory use. While less pre-cise at elevated levels, ChatGPT produces more concise and idiomatic responses. Both models had limited competence at the expert level; however, DeepSeek-R1 indicated a slight edge. The study illustrates a trade-off between accuracy and code clarity, so as to inform the selection of LLMs based on task requirements and provide a foundation for future efforts in optimizing code generation models for actual applications.